[RFC 01/12] exynos-fimc-is: Adding device tree nodes

[Sylwester Nawrocki]

Hi Arun,

On 03/28/2013 06:10 AM, Arun Kumar K wrote:
> On Wed, Mar 27, 2013 at 7:17 PM, Sylwester Nawrocki
> <s.nawrocki at samsung.com>  wrote:
>> On 03/27/2013 05:31 AM, Arun Kumar K wrote:
>>> On Wed, Mar 27, 2013 at 4:21 AM, Sylwester Nawrocki
>>> <sylvester.nawrocki at gmail.com>  wrote:
>>>> On 03/26/2013 01:17 PM, Arun Kumar K wrote:
>> [...]
>>> Only issue is with the context sharing.
>>> Right now you can see that the fimc-is context is shared between all
>>> the subdevs.
>>> As all of them are part of the same driver, it is possible.
>>> If sensor is made as an independent i2c device, a separate probe will
>>> be called for it.
>>> For ISP sensor subdev to work independently, it needs to call the
>>> fimc_is_pipeline_* calls
>>> for FW initialization and other configurations for which it needs the
>>> fimc-is main context.
>>>
>>> Now there is a workaround here by calling a get_context() macro in
>>> sensor's probe
>>> to get the fimc-is context. This will cause the same context to be
>>> shared and updated by
>>> 2 drivers though both are part of fimc-is.
>>> Is this acceptable? Or am I missing some other simple solution of implementing
>>> it in a better way?
>>
>> OK, thanks for the explanation.
>>
>> I can think of at least one possible way to get hold of the fimc-is
>> context in the subdev. For instance, in subdev's .registered callback
>> you get a pointer to struct v4l2_device, which is normally embedded
>> in a top level driver's private data. Then with container_of()
>> you could get hold of required data at the fimc-is driver.
>
> But as per current implementation, it is not the fimc-is driver that is
> registering the ISP subdevs. It will be registered from the
> media controller driver. So fimc-is context cannot be obtained by
> just using container_of().

I guess best option would be to have a function to get the IS slave
interface driver context at the sensor subdev exported by the IS driver
module, as you suggested previously.

You still could obtain the fimc-is object from the media device private
data structure, since the media device has normally a list of its all
entities in one form or the other. But the sensor would need to know
details of the media device, which makes it a bit pointless.

Nevertheless, my main concern is the DT binding. Sharing the sensor
subdev driver might not be that important at the moment, we are talking
about 300..500 lines of code per ISP driver currently anyway.

More important is to have the hardware described in a standard way, so
when the firmware changes there is no need to change the DT bindings.

>> But... to make the subdev drivers reuse possible subdevs should
>> normally not be required to know the internals of a host driver they
>> are registered to. And it looks a bit unusual to have fimc_pipeline_*
>> calls in the sensor's operation handlers.
>
> fimc_pipeline_* I mentioned is not the media controller pipeline.

Ok, I admit I got confused a bit, since the word "pipeline" refers in the
code to both: the internal ISP chain, and the data processing chain
containing the FIMC-IS and other devices like CSI-2 receiver or GScaler.

> In the fimc-is driver, all the subdevs just implement the interface part.
> All the core functionalities happen in fimc-is-pipeline.c and
> fimc-is-interface.c.
> Since these ISP subdevs (sensor, isp, scc, scp) are not independent
> devices, all are controlled by the ISP firmware whose configuration and
> interface is done from the fimc_is_pipeline_* and fimc_is_itf_* functions.
> So all the ISP subdevs including sensor need to call these functions.
>
>>
>> I thought that the subdevs could provide basic methods and it would
>> be the top level media driver that would resolve any dependencies
>> in calling required subdev ops, according to media graph configuration
>> done by the user on /dev/media?.
>>
>
> In case of ISP subdevs (isp, scc and scp), there is not much configuration
> that the media device can do. Only control possible is to turn on/off
> specific scaler DMA outputs which can be done via the video node ioctls.
> The role of media device here is mostly to convey the pipeline structure
> to the user. For eg. it is not possible to directly connect isp (sd)
> -->  scp (sd).
> In the media controller pipeline1 implementation, we were planning to
> put immutable links between these subdevs. Is that acceptable?

Not sure I understand which links you mean exactly. Could you post the
media graph generated by media-ctl (--print-dot) ?

If you're talking about the on-the-fly (FIFO) links, then it probably
makes sense. The media device driver should respond to the link_notify
events and not to allow data links unsupported in the hardware. If you
create immutable OTF links, then how would you switch between DMA and
OTF interfaces ? Or can all processing blocks of the ISP chain work
simultaneously with the DMA and OTF ? The FD block, for instance, can fed
data from memory _or_ from previous processing block in the chain, right ?
You will need a user interface to control which input is used and the
links configuration seems most natural here.

>> The media driver has a list of media entities (subdevices and video
>> nodes) and I though it could coordinate any requests involving whole
>> video/image processing pipeline originating from /dev/video ioctls/fops.
>>
>> So for instance if /dev/video in this pipeline is opened
>>
>> sensor (sd) ->  mipi-csis (sd) ->  fimc-lite (sd) ->  memory (/dev/video)
>>
>> it would call s_power operation on the above subdevs and additionally
>> on e.g. the isp subdev (or any other we choose as a main subdev
>> implementing the FIMC-IS slave interface).
>>
>> Then couldn't it be done that video node ioctls invoke pipeline
>> operations, and the media device resolves any dependencies/calls
>> order, as in case of the exynos4 driver ?
>
> On Exynos4 subdevs, it is well and good since all the subdevs are
> independent IPs. Here in ISP since the same IP can take one input and

Not really, there are going to be 2 subdevs exposed by the fimc-is: ISP
and FD. However FD is still not supported in my last patch series. I was
planning this for a subsequent kernel release.

> provide multiple outputs, we designed them as separate subdevs. So
> here we cannot make the subdevs independent of each other where only
> the sequence / dependencies is controlled from the media device.

I'm not asking you to make the FIMC-IS subdevs more self-contained,
it's of course perfectly fine to have multiple (logical) subdevs exposed
by a complex device like that. I have been thinking only about the
sensor driver, since the sensors are normally shared across ISPs from
various chip manufacturers. But let us leave this topic for now.

BTW, in my interpretation FIMC-IS is a collection of IPs/peripheral
devices, not a single black box, including an image sensor. And the
firmware should not be the most significant factor how we expose
the whole subsystem to the user. All elements of the ISP chain, i.e.
an IP control registers are visible to both, the main CPU and the
FIMC-IS (Cortex-A5) MCU. Still, it would be possible to create v4l2
subdevs dynamically, depending on the firmware architecture.

---

Regards,
Sylwester